CN111272883A - An intelligent detection and identification method of rock fracture mode based on acoustic emission model - Google Patents

Abstract

The invention discloses an intelligent detection and identification method of rock rupture mode based on acoustic emission model, comprising the following steps: firstly, an acoustic emission system for testing acoustic emission parameters of rock rupture process is arranged on the rock to be monitored, and then target characteristic data is input The pre-trained signal recognition model is pre-trained through the training set of rock rupture acoustic emission, and then intelligently identifies the ratio of tension and shear crack development during the rock rupture process, and finally determined according to the rock rupture acoustic emission signal. There is a corresponding relationship between waveform characteristics and rock fracture pattern recognition, which provides a series of reliable detection thresholds for quantitatively formulating rock mass disaster early warning schemes, and provides an analysis method for in-depth research and identification of rock fracture and instability precursor signal characteristics.

Description

An intelligent detection and identification method of rock fracture mode based on acoustic emission model

technical field

The invention relates to the technical field of geological survey applications, in particular to an intelligent detection and identification method for rock fracture modes based on an acoustic emission model.

Background technique

Acoustic emission signal detection technology provides an attractive solution for damage assessment/structural health monitoring of various rock structures (slopes, dams, subgrades, tunnels, etc.). The performance and function of these civil structures are relevant to the safety of society, and during various natural events (ie earthquakes, hurricanes and tsunamis), these events may compromise their safety and availability. To ensure the overall stability of these structures, the correct assessment and prediction of the development of rock fractures is crucial, especially in engineering practice, because models of rock fractures reflect not only their condition as a material, but also the entire The state of the system at the structural level.

Acoustic emission (AE)-based methods offer an attractive solution for crack nucleation and propagation in rock structures. The present invention proposes that the intelligent detection and identification method of rock fracture mode based on Gaussian Mixture Model (GMM) is a distribution-based unsupervised classification technology, which has been successfully applied in many fields, including sound recognition, image processing, dynamic system and tracking and text recognition; however, this technology has not been used for intelligent recognition of rock fracture patterns based on acoustic emission. Based on the above reasons, a probability scheme for rock fracture mode classification based on acoustic emission Gaussian mixture model is proposed.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a method for intelligent detection and identification of rock fracture modes based on acoustic emission model in order to solve the shortcomings existing in the prior art.

In order to realize the purpose of the present invention, the present invention proposes the following technical solutions:

An intelligent detection and identification method of rock rupture mode based on acoustic emission model, comprising the following steps:

Step 1, arranging an acoustic emission system for testing the acoustic emission parameters of the rock fracture process on the rock to be monitored;

In step 2, the acoustic emission parameters collected by the acoustic emission system during the rock fracture process are input into the pre-trained signal identification model, and the signal identification model is pre-trained through the training set of rock fracture acoustic emission;

Step 3, the signal identification model intelligently identifies the development ratio of tension and shear cracks in the process of rock fracture;

Step 4: According to the relationship between the waveform characteristics determined by the rock fracture acoustic emission signal and the rock fracture pattern recognition, a series of reliable detection thresholds are provided for quantitatively formulating the rock mass disaster early warning plan, and at the same time, the characteristics of the rock fracture and instability precursor signals are identified for in-depth research. Provide an analysis method.

Preferably, in step 1, the acoustic emission system selects and measures the ringing count, duration, peak frequency and rise time in the rock acoustic signal to analyze the process of rock fracture.

Preferably, in step 1, the method for the acoustic emission system to collect rock acoustic signals is based on the JCMS parameter analysis method:

The average frequency _AF of acoustic emission parameters is obtained by ringing count/duration, and _RA is obtained by rise time/peak amplitude, and then the two groups of data are classified.

Preferably, in step 2, the preset training set of the signal identification model includes a Gaussian Mixture Model (Gaussian Mixture Model, GMM for short) and an Expectation Maximization (Expectation Maximization, EM for short) algorithm.

Preferably, in step 2, according to the relationship between _AF and _RA , the analysis of tension and shear cracks is carried out, and when the analysis of tension and shear cracks is carried out, the mixed Gaussian model and the expected maximum algorithm are combined. As a training model, by observing the closeness of the sampled probability value and the model probability value, it is judged whether a model is fit and good, and the relationship between _AF and _RA is intelligently detected and identified.

Preferably, in step 2, by adjusting the model to make the new model more suitable for the probability value, iterate this process multiple times until the two probability values are very close, stop updating and complete the model training, and use this process with algorithm to achieve:

The expected value of the data is calculated by the Gaussian mixture model. The Gaussian mixture model itself is a parametric probability density function, expressed as the weight of the Gaussian density of the M components, and the mean μ and standard deviation σ of the distribution are updated through continuous iteration to maximize the expected value until These two parameters change very little;

For D-dimensional measurement and training, the mixing density is defined as:

为单模态的高斯(正常)密度，

为特征向量；In the formula, ω _i is the mixed weight,

is the single-modal Gaussian (normal) density,

is the feature vector;

The Gaussian component density of each single mode is in the form of a D-variable Gaussian function as:

为D×1的平均向量，∑i为D×D的协方差矩阵；In the formula,

is the average vector of D×1, ∑i is the covariance matrix of D×D;

完整的混合高斯模型应由平均向量

协方差矩阵In order to make the mixed weight ω _i satisfy

The full Gaussian mixture model should consist of the mean vector

covariance matrix

The mixed weighting of ∑i and all component densities M is used to parameterize λ, and the parameter λ is expressed by equation (3) as:

的分布匹配，确定λ的最佳估值；For a classification system based on a mixture of Gaussian models, the goal of model training is to estimate the λ of the mixture Gaussian model parameters such that the Gaussian mixture density is related to the eigenvectors

The distribution matches to determine the best estimate of λ;

和∑i的常用方法之一，最大似然值估计估计能在给定训练数据的情况下使混合高斯模型的可能性最大化，对于一系列T训练向量

假定各向量之间是独立的，可以写成The maximum likelihood value estimation (Maximum Likelihood, ML for short) is used to estimate ω _i ,

One of the common methods of and ∑i, maximum likelihood estimation estimates the possibility of maximizing the likelihood of a Gaussian mixture model given training data, for a series of T training vectors

Assuming that each vector is independent, it can be written as

Since this expression is a nonlinear function of λ, it is computationally intractable to directly maximize (that is, setting the first derivative equal to zero and constraining the second derivative to be positive), so consider using the Expectation-maximization algorithm (EM) Iterate to get ML parameters.

时，给定第 k个重新估计的模型λ^k Preferably, in step S2, the training process of the expectation-maximization algorithm is an iterative process, starting from the initial model λ ^k and then estimating a new model λ ^k+1 , so that p(X|λ ^k+1 )>p(X|λ ^k ), so the new model becomes the initial model for the next iteration, and this process is repeated until a certain convergence threshold is reached (such as the log-likelihood value of 1026), the algorithm is determined by the expectation and maximization consists of two steps, which guarantees a monotonically increasing model relief value, the result of the expectation step is the posterior probability for the i-th component, which is defined as the probability of state i, when the m-th Gaussian mixture results for

, given the k-th re-estimated model λ ^k

分别由式(6)(7)(8)用最大化步骤来返回分布参数：In the formula,

The distribution parameters are returned by the maximization step by equations (6) (7) and (8) respectively:

(或测量向量)认为是一个二维向量(即

)，当有T个训练向量时序列

两种分类对应张拉和剪切模式分别是I＝{1，2}，此时再“估计”混合高斯模型的参数(每个隐藏类的权重，均值和协方差矩阵)，使它们与训练特征向量

的分布最为匹配。In this way, the mixed Gaussian model can classify structures with two types of crack modes, such as rock and concrete, namely tension and shear crack classification (M=2). In order to classify these two crack modes, the feature vector

(or measurement vector) is considered a two-dimensional vector (i.e.

), when there are T training vectors, the sequence

The corresponding tension and shear modes of the two classifications are I = {1, 2}. At this time, the parameters of the Gaussian mixture model (weight, mean and covariance matrix of each hidden class) are "estimated" to make them consistent with the training Feature vector

distribution that best matches.

Compared with the prior art, the beneficial effects of the present invention are:

The present invention proposes an intelligent detection and identification method of rock fracture mode based on acoustic emission model, provides a solution for the disaster types of sudden brittle fracture of rock mass such as rock collapse and rock landslide, and breaks through the traditional indirect monitoring of deformation of rock. It solves the objective limitations of "poor real-time, insufficient precursors, and low success rate" for early warning of rock mass damage and damage, and solves the effective monitoring and early warning technology approach for sudden rock mass brittle instability and damage disasters. , rockfall, landslide) disaster prevention and mitigation and emergency disaster relief provides effective scientific and technological support, with very important scientific significance and application value.

Description of drawings

Other features, objects and advantages of the present invention will become more apparent by reading the detailed description of non-limiting embodiments with reference to the following drawings:

Figure 1 is a classification diagram of rock cracks;

Figure 2 shows the crack identification results of limestone under uniaxial compression at the initial and middle and late stages of stress _σc ;

Figure 3 is a list of percentages of limestone tension and shear crack stress intervals;

Figure 4 shows the proportions of the two types of cracks in limestone at each loading stage.

Detailed ways

The technical solutions in the embodiments of the present invention will be described clearly and completely below. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments.

An intelligent detection and identification method of rock rupture mode based on acoustic emission model, which specifically includes the following steps:

Step 1, arranging an acoustic emission system for testing the acoustic emission parameters of the rock fracture process on the rock to be monitored;

In step 1, the acoustic emission system selects the ringing count, duration, peak frequency and rise time in the rock acoustic signal to analyze the process of rock fracture.

In step 1, the method of the acoustic emission system to collect the rock acoustic signal is based on the JCMS parameter analysis method:

The average frequency _AF of acoustic emission parameters is obtained by ringing count/duration, and _RA is obtained by rise time/peak amplitude, and then the two groups of data are classified.

Step 2, input the target feature data into the pre-trained signal identification model, and the signal identification model is obtained in advance through the training set of rock rupture acoustic emission;

In step 2, the preset training set of the signal identification model includes a Gaussian Mixture Model (Gaussian Mixture Model, GMM for short) and an Expectation Maximization (Expectation Maximization, EM for short) algorithm.

In step 2, according to the relationship between _AF and _RA , the analysis of tension and shear cracks is carried out. When the analysis of tension and shear cracks is carried out, the mixture Gaussian model and the expectation-maximization algorithm are combined as the training model. By observing the closeness of the sampled probability value and the model probability value, it is judged whether a model is fit and good, and the relationship between _AF and _RA is intelligently detected and identified.

In step 2, by adjusting the model to make the new model more suitable for the probability value, iterate this process many times until the two probability values are very close, stop the update and complete the model training, and implement this process with an algorithm:

The expected value of the data is calculated by the Gaussian mixture model. The Gaussian mixture model itself is a parametric probability density function, expressed as the weight of the Gaussian density of the M components, and the mean μ and standard deviation σ of the distribution are updated through continuous iteration to maximize the expected value until These two parameters change very little;

For D-dimensional measurement and training, the mixing density is defined as:

为单模态的高斯(正常)密度，

为特征向量；In the formula, ω _i is the mixed weight,

is the single-modal Gaussian (normal) density,

is the feature vector;

The Gaussian component density of each single mode is in the form of a D-variable Gaussian function as:

为D×1的平均向量，∑i为D×D的协方差矩阵；In the formula,

is the average vector of D×1, and ∑i is the covariance matrix of D×D;

完整的混合高斯模型应由平均向量

协方差矩阵In order to make the mixed weight ω _i satisfy

The full Gaussian mixture model should consist of the mean vector

covariance matrix

The mixed weighting of ∑i and all component densities M is used to parameterize λ, and the parameter λ is expressed by equation (3) as:

的分布匹配，确定λ的最佳估值；For a classification system based on a mixture of Gaussian models, the goal of model training is to estimate the λ of the mixture Gaussian model parameters such that the Gaussian mixture density is related to the eigenvectors

The distribution matches to determine the best estimate of λ;

和∑i的常用方法之一，最大似然值估计估计能在给定训练数据的情况下使混合高斯模型的可能性最大化，对于一系列T训练向量

假定各向量之间是独立的，可以写成The maximum likelihood value estimation (Maximum Likelihood, ML for short) is used to estimate ω _i ,

One of the common methods of and ∑i, maximum likelihood estimation estimates the possibility of maximizing the likelihood of a Gaussian mixture model given training data, for a series of T training vectors

Assuming that each vector is independent, it can be written as

Since this expression is a nonlinear function of λ, it is computationally intractable to directly maximize (that is, setting the first derivative equal to zero and constraining the second derivative to be positive), so consider using the Expectation-maximization algorithm (EM) Iterate to get ML parameters.

时，给定第k个重新估计的模型λ^k In step S2, the training process of the expectation-maximization algorithm is an iterative process, starting from the initial model λ ^k and then estimating a new model λ ^k+1 , so that p(X|λ ^k+1 )>p( ^X |λk ), so that the new model becomes the initial model for the next iteration, and the process is repeated until a certain convergence threshold is reached (such as the log-likelihood value of 1026), the algorithm consists of the expectation and maximization of two The result of the expected step is the posterior probability for the i-th component, which is defined as the probability of state i, when the m-th Gaussian mixture results in

, given the k-th re-estimated model λ ^k

分别由式(6)(7)(8)用最大化步骤来返回分布参数：In the formula,

The distribution parameters are returned by the maximization step by equations (6) (7) and (8) respectively:

(或测量向量)认为是一个二维向量(即

)，当有T个训练向量时序列

两种分类对应张拉和剪切模式分别是I＝{1，2}，此时再“估计”GMM的参数(每个隐藏类的权重，均值和协方差矩阵)，使它们与训练特征向量

的分布最为匹配。In this way, GMM can classify structures with two types of crack modes, such as rock and concrete, namely tension and shear crack classification (M=2). In order to classify these two crack modes, the feature vector

(or measurement vector) is considered a two-dimensional vector (i.e.

), when there are T training vectors, the sequence

The corresponding tension and shear modes of the two classifications are I = {1, 2}, at which time the parameters of the GMM (weight, mean and covariance matrix of each hidden class) are "estimated" so that they are the same as the training feature vector.

distribution that best matches.

Step 3, intelligently identify the ratio of tension and shear crack development during rock fracture, as shown in Figure 1;

Step 4: There is a corresponding relationship between the waveform characteristics determined by the acoustic emission signal of rock rupture and the recognition of rock rupture patterns, providing a series of reliable detection thresholds for quantitatively formulating rock mass disaster early warning plans, and at the same time identifying rock rupture and instability precursor signals for in-depth research. Features provide a method of analysis.

Example 1

The whole training can be summarized as:

① Initialize the parameters in λ, and use the vector quantization method to preliminarily determine the two encoded parameters of the Gaussian mixture under the state correlation;

②Apply formula (5) to get Pr(i|x _t , λ ^k );

③Use Pr(i|x _t , λ ^k ) to better estimate the parameter λ ^k+1 (see equations (6) to (8));

④ Iterate steps ② and ③ until convergence.

As shown in Fig. 2, (a) and (b) are the intelligent crack identification results of limestone under uniaxial compression at the initial and middle and late stages of stress _σc , respectively. It can be observed from the figure that at the initial stage of loading (0～0.1) _σc of limestone is almost all tensile cracks, the ellipse of the tensile cluster is relatively round, and the points around the center point are scattered evenly. 0.6) σ _c develops into the transition stage from tensile to shear, at this time, under the action of larger load, _AF usually has a higher amplitude, so the value of _RA changes in a small range, and the clustering is high. The probability region gradually moves to the mean range of the shear class, and the two mutually exclusive categories (shear and stretch) begin to gradually form a confluence (mixture), but these two categories are still divided. At this stage, the high probability region is almost The ensemble is concentrated around the mean of the clipping class.

This patent has been well used in indoor uniaxial compression and acoustic emission tests. Figure 3 shows the percentage of limestone tension cracks and shear cracks in the entire loading stage. The proportion of % shear cracks reaches the maximum value, and the rock sample has entered the later stage of the unstable growth stage. In this study, the maximum proportion of limestone shear cracks is 44.59%, and this value is used as a precursor threshold for predicting limestone failure. When the percentage exceeds this threshold, an early warning can be triggered as a serious damage to the rock mass. prediction. At the same time, the values of _RA and _AF at the center of the cluster of two types of failure cracks have the characteristics of acoustic emission signals with low _AF and high _RA values for shear cracks, and the characteristics of high _AF and low _RA values for tensile cracks. This is the same as the _RA and AF values of the tensile and shear cracks obtained by the _JCMS parametric analysis method.

Finally, to verify the classification results of cracks for all loading steps, Figure 4 shows the proportion of AE activity associated with two types of crack clusters in limestone at each loading step. It can be found that tension cracks play a dominant role in the whole loading process of limestone, that is, most of the acoustic emission signals are generated by the nucleation of tension cracks. Unlike the reinforced concrete four-point bending test, the crack failure mode of the rock chamber acoustic emission test can be clearly divided into three stages: (1) the dominant action stage of tension, the initial loading step, and the eigenvectors are concentrated near the average value of the tension class; ② Transition stage, the transition stage from tensioning to shearing in the intermediate loading step, in this stage, the high probability region gradually moves to the mean value of the shear class; ③ Damage stage, during the final loading step, it is shear crack control, which is At this stage, the areas most likely to fail are concentrated around the average value of shear cracks. The reason is that the two loading methods are different and the material uniformity is different. Although the proportion of the two types of cracks in the whole loading stage of the rock chamber acoustic emission test has no obvious regularity, we can still find that the maximum proportion of the shear crack occurs in the stage of 80% to 90% of the total loading time. This time period corresponds to the middle and late stages of the unstable expansion stage in the rock loading process, which is used as a precursor to failure.

The above description is only a preferred embodiment of the present invention, but the protection scope of the present invention is not limited to this. The equivalent replacement or change of the inventive concept thereof shall be included within the protection scope of the present invention.

Claims (7)

1. An intelligent detection and identification method for a rock fracture mode based on an acoustic emission model is characterized by comprising the following steps:

step 1, arranging an acoustic emission system for testing acoustic emission parameters in a rock breaking process on a rock to be monitored;

step 2, inputting target characteristic data into a pre-trained signal recognition model, wherein the signal recognition model is obtained by training a training set of rock fracture acoustic emission in advance;

step 3, intelligently identifying the proportion of tension and shear crack development in the rock cracking process;

and 4, according to the corresponding relation between the waveform characteristics determined by the rock cracking acoustic emission signals and the rock cracking mode identification, providing a series of reliable detection threshold values for quantitatively formulating a rock disaster early warning scheme, and simultaneously providing an analysis method for deeply researching and identifying rock cracking instability precursor signal characteristics.

2. The intelligent detection and identification method for rock fracture patterns based on acoustic emission model is characterized in that in step 1, the acoustic emission system selects and measures the ringing count, duration, peak frequency and rise time in the rock acoustic signal to analyze the rock fracture process.

3. The method for intelligent detection and identification of rock cracking patterns based on acoustic emission model as claimed in claim 2, wherein in step 1, the method for collecting rock acoustic signals by the acoustic emission system is based on JCMS parameter analysis, i.e. the average frequency A of acoustic emission parameters is obtained by dividing the ringing count by the duration_FDividing the rise time by the peak amplitude to obtain R_AAnd then, the two groups of data are classified.

4. The method as claimed in claim 3, wherein in step 2, the preset training set of the signal recognition Model includes Gaussian Mixture Model (GMM) and Expectation Maximization (EM) algorithm.

5. The intelligent detection and identification method for rock rupture modes based on acoustic emission model as claimed in claim 4, characterized in that in step 2, according to A_FAnd R_AThe relation between the model and the model is used for analyzing tension and shear cracks, when tension and shear crack analysis is carried out, a Gaussian mixture model and an expected maximum algorithm are combined to serve as training models, the probability value of sampling and the closeness degree of the probability value of the model are observed to judge whether the model is fit or good, and the A is compared with the A_FAnd R_AThe relationship between them is intelligently detected and identified.

6. The intelligent detection and recognition method for rock rupture modes based on acoustic emission model as claimed in claim 5, wherein in step 2, the process is iterated for a plurality of times by adjusting the signal recognition model to make the new signal recognition model more adaptive to the probability value, and the updating is stopped and the model training is completed until the two probability values are very close, and the process is implemented by an algorithm:

calculating expected values of data through a Gaussian mixture model, wherein the Gaussian mixture model is a parameter probability density function and is expressed as weighting of Gaussian density of M components, and the expected values are maximized by continuously iterating to update the mean value mu and the standard deviation sigma of distribution until the two parameters change very little;

for D-dimensional measurement, training, the mixture density is defined as:

in the formula, ω_iIn order to mix the weight values, the user can select the weight value,

is a single mode gaussian (normal) density,

is a feature vector;

the gaussian component density of each single mode is in the form of a D-variant gaussian function:

in the formula (I), the compound is shown in the specification,

the vector is an average vector of Dx1, and the sigma i is a covariance matrix of DxD;

to let the mixed weight omega_iSatisfy the requirement of

The complete Gaussian mixture model should be formed by averaging vectors

The covariance matrix Σ i and the mixed weighting of all component densities M parameterize it λ, which is expressed by equation (3):

for a classification system based on a Gaussian mixture model, the goal of model training is to estimate the lambda of the parameters of the Gaussian mixture model so that the Gaussian mixture density and the feature vector

Determining the optimal estimate of λ;

maximum Likelihood estimation (ML) is used to estimate ω_i、

And Σ i, the maximum likelihood estimation estimate maximizes the probability of a gaussian mixture model given the training data, for a series of T training vectors

Given the independence between vectors, can be written as

This expression, as a non-linear function of λ, is computationally intractable to directly maximize (i.e. set the first derivative equal to zero and constrain the second derivative to be positive), considering the ML parameters obtained by iteration through the Expectation-maximization algorithm (EM for short).

7. The method for intelligent detection and identification of rock cracking patterns based on acoustic emission model as claimed in claim 6, wherein in step S2, the training process of expectation-maximization algorithm is an iterative process from the initial model λ^kInitially, a new model λ is then estimated^k+1Thus, there is p (X | λ)^k+1)＞p(X|λ^k) So that the new model becomes the initial model for the next iteration and the process is repeated until a certain convergence threshold is reached (e.g. the log likelihood is 1026), the algorithm consisting of two steps of expectation and maximization, which ensures a monotonic increase in the model's belief value, the result of the expectation step being the posterior probability for the i-th component, which is defined as the probability of state i, when the m-th mixture of gaussians results in state i

Given the kth re-estimated model λ^k

In the formula, ω_i′

The distribution parameters are returned by equations (6), (7) and (8), respectively, with a maximization step:

the gaussian mixture model can classify structures having two types of crack patterns, i.e., tensile crack and shear crack (M is 2), such as rock and concrete, and the feature vector is used to classify the two types of crack patterns

(or measurement vector) is considered to be a two-dimensional vector (i.e., a vector of measurements)

) When there are T training vectors, the sequence

The two classes correspond to the pull and shear modes I ═ 1, 2, respectively, at which point the parameters of the gaussian mixture model (weight, mean and covariance matrices for each hidden class) are "estimated" and matched to the training eigenvectors

Are most closely matched.

Images